Luke’s Dev Lab

Overview

A repository for development logs, model evaluations, and technical notes curated by Luke’s Dev Lab.

Recent Evaluations

Swift 1.5 Qwen3.8-27B GSQ-RCO IQ3_S 16GB LLM Performance Benchmark

Date: 2026-10-06 Focus: Quantization, Local Deployment, 16GB VRAM Optimization

• Model: UkisAI Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF
• Quantization: IQ3_S
• Hardware: RTX 2000 Ada (16GB VRAM) on Ubuntu server
• Key Findings:
  • Comprehensive benchmark of IQ3_S quantization performance.
  • Validated viability for local LLM setups with constrained VRAM.
  • Swift 1.5 Qwen3.8-27B GSQ-RCO IQ3_S 16GB LLM Performance Benchmark
• Source: Swift 1.5 Qwen3.8-27B GSQ-RCO IQ3_S 16GB LLM Performance Benchmark

Qwen 3.8-Max Performance Benchmarks and Agentic Task Evaluation

Date: 2026-09-25 Focus: Performance, Agentic Tasks, ClinePass Integration

• Model: Qwen 3.8-Max
• Context: Evaluated via David AU variant workflows using ClinePass.
• Key Findings:
  • Demonstrated excellent prefill speeds, ranging from approximately 640 tokens/second for short prompts.
  • Assessment of agentic task capabilities beyond standard benchmarking.
  • Qwen 3.8-Max Performance Benchmarks and Agentic Task Evaluation

Qwen3.8 27B T

Source Notes

• 2026-10-06: Swift 1.5 Qwen3.8-27B GSQ-RCO IQ3_S 16GB LLM Performance Benchmark · ▶ source
• 2026-09-26: Q2 Re-evaluation: Benchmarking Performance, Memory, Reasoning · ▶ source
• 2026-09-25: Qwen 3.8-Max Performance Benchmarks and Agentic Task Evaluation · ▶ source
• 2026-09-23: Qwen3.8 27B Turbo Fable Cold Fusion LLM: 16GB Local Performance Benchmark · ▶ source
• 2026-09-22: Ternary Bonsai 2 27B GGUF 1-bit 2-bit Performance, Memory, Reasoning Evaluation · ▶ source
• 2026-09-12: Nail-Qwen 35B A3B LLM: Performance, Reasoning, Coding on 16GB GPU Evaluation · ▶ source